Cmprsr: Abstractive Token-Level Question-Agnostic Prompt Compressor
Published in arXiv preprint — under review at ACL 2026, 2025
Summary
Most prompt compressors are query-specific: they need the question up front in order to decide what to keep. That is fine for a single-turn QA pipeline and awkward everywhere else — caches, shared context, and any setting where the same context serves many different questions.
Cmprsr is question-agnostic. It compresses at the token level without seeing the query, so one compressed prompt can serve whatever gets asked of it afterwards. It is trained with supervised fine-tuning followed by preference optimisation.
Results
At matched token budgets, Cmprsr beats LLMLingua-2 on:
- mathematical reasoning
- summarisation
- multiple-choice question answering
Why It Mattered Afterwards
This line of work started at EPFL’s dlab as a benchmarking programme measuring LLMs as compressors, and became the research basis for a production compression API — the one now shipped by Compresr.
Publication Details
- Preprint: arXiv:2511.12281
- Status: under review at ACL 2026
- My role: co-author